Collaborative Laboratory Edge Networks
Abstract
Chemistry language models now predict reactions and retrosynthetic routes with useful accuracy, but they assume that proprietary reaction records can be pooled into a central cluster and that inference runs on abundant accelerators. Neither assumption holds in pharmaceutical and fine-chemical practice, where yield records cannot leave the plant network, the points of use are heterogeneous resource-limited sites, and different laboratories occupy disjoint regions of chemical space. We present ChemFedFusion, a heterogeneity-aware framework for federated chemistry-model training and offloading over collaborative laboratory edge networks. ChemUCP is a chemistry-aware utility–cost predictor that scores a task–node pair by expected chemistry gain, expected latency, and an explicit out-of-distribution risk of emitting unparsable or unreliable chemistry, using a dual-channel tokenizer and a scaffold-coverage encoder that is corrected online from execution feedback. CSAW replaces sample-count aggregation with the product of a scaffold-space complementarity term, a site reliability term, and a weakened volume guard, amplifying rare chemistry while suppressing noisy sites. ARCO couples offloading and aggregation through a shared adaptation matrix and an event-driven asynchronous aggregator with a staleness bound, so that successful offloads teach the aggregator which node is competent for which chemistry task. On Mol-Instruction and USPTO-50K, ChemFedFusion reaches 0.883 Exact and 51.36% Top-1, and on the FedChem suite it reduces regression RMSE by up to 9.99% and classification ROC-AUC improvements are consistent across all nine datasets and three heterogeneity levels. The framework reduces convergence rounds by 22.4% and raises the fraction of answers that are both on time and chemically valid by 8.8 percentage points. A failure analysis shows that the remaining errors are dominated by synthetically infeasible routes that parse correctly, which no validity-based router can detect.
Related
- Human–Robot Collaborative Reaction Development with a Self-Driving Laboratory
- Projective Geometric Algebra Graph Neural Networks: Edge-Modulated Message Passing via the Geometric Product
- Large-Scale Collaborative Assessment of Binding Free Energy Calculations for Drug Discovery Using OpenFE
- Enhancing Boiler Operational Economy via Data-Driven Collaborative Combustion Optimization
- Goggles On, Bagels Out: A Randomized Hazard Hunt in a Browser Replica of a Real Teaching Laboratory
BibTeX
@misc{williams2026collaborative,
title = {{Collaborative Laboratory Edge Networks}},
author = {Andre Williams and Shanice Brown and Kevin Thompson and Jamar Campbell and Alicia Morgan},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15009129/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009129/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Collaborative Laboratory Edge Networks
AU - Andre Williams
AU - Shanice Brown
AU - Kevin Thompson
AU - Jamar Campbell
AU - Alicia Morgan
PY - 2026
DA - 2026/09/21
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009129/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009129/v1
AB - Chemistry language models now predict reactions and retrosynthetic routes with useful accuracy, but they assume that proprietary reaction records can be pooled into a central cluster and that inference runs on abundant accelerators. Neither assumption holds in pharmaceutical and fine-chemical practice, where yield records cannot leave the plant network, the points of use are heterogeneous resource-limited sites, and different laboratories occupy disjoint regions of chemical space. We present ChemFedFusion, a heterogeneity-aware framework for federated chemistry-model training and offloading over collaborative laboratory edge networks. ChemUCP is a chemistry-aware utility–cost predictor that scores a task–node pair by expected chemistry gain, expected latency, and an explicit out-of-distribution risk of emitting unparsable or unreliable chemistry, using a dual-channel tokenizer and a scaffold-coverage encoder that is corrected online from execution feedback. CSAW replaces sample-count aggregation with the product of a scaffold-space complementarity term, a site reliability term, and a weakened volume guard, amplifying rare chemistry while suppressing noisy sites. ARCO couples offloading and aggregation through a shared adaptation matrix and an event-driven asynchronous aggregator with a staleness bound, so that successful offloads teach the aggregator which node is competent for which chemistry task. On Mol-Instruction and USPTO-50K, ChemFedFusion reaches 0.883 Exact and 51.36% Top-1, and on the FedChem suite it reduces regression RMSE by up to 9.99% and classification ROC-AUC improvements are consistent across all nine datasets and three heterogeneity levels. The framework reduces convergence rounds by 22.4% and raises the fraction of answers that are both on time and chemically valid by 8.8 percentage points. A failure analysis shows that the remaining errors are dominated by synthetically infeasible routes that parse correctly, which no validity-based router can detect.
ER -